# Copyright 2025 The HuggingFace Datasets Authors and the current dataset script contributor. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. """BirdSet: A Large-Scale Dataset for Audio Classification in Avian Bioacoustics""" import os import datasets import pandas as pd from tqdm.auto import tqdm import tarfile from pathlib import Path from . import classes from .classes import BIRD_NAMES_NIPS4BPLUS, BIRD_NAMES_AMAZON_BASIN, BIRD_NAMES_HAWAII, \ BIRD_NAMES_HIGH_SIERRAS, BIRD_NAMES_SIERRA_NEVADA, BIRD_NAMES_POWDERMILL_NATURE, BIRD_NAMES_SAPSUCKER, \ BIRD_NAMES_COLUMBIA_COSTA_RICA, BIRD_NAMES_XENOCANTO, BIRD_NAMES_XENOCANTO_M from .descriptions import _NIPS4BPLUS_CITATION, _NIPS4BPLUS_DESCRIPTION, \ _HIGH_SIERRAS_DESCRIPTION, _HIGH_SIERRAS_CITATION, _SIERRA_NEVADA_DESCRIPTION, _SIERRA_NEVADA_CITATION, \ _POWDERMILL_NATURE_DESCRIPTION, _POWDERMILL_NATURE_CITATION, _AMAZON_BASIN_DESCRIPTION, _AMAZON_BASIN_CITATION, \ _SAPSUCKER_WOODS_DESCRIPTION, _SAPSUCKER_WOODS_CITATION, _COLUMBIA_COSTA_RICA_CITATION, \ _COLUMBIA_COSTA_RICA_DESCRIPTION, _HAWAIIAN_ISLANDS_CITATION, _HAWAIIAN_ISLANDS_DESCRIPTION ############################################# _BIRDSET_CITATION = """\ @misc{rauch2025birdsetlargescaledatasetaudio, title={BirdSet: A Large-Scale Dataset for Audio Classification in Avian Bioacoustics}, author={Lukas Rauch and Raphael Schwinger and Moritz Wirth and René Heinrich and Denis Huseljic and Marek Herde and Jonas Lange and Stefan Kahl and Bernhard Sick and Sven Tomforde and Christoph Scholz}, year={2025}, eprint={2403.10380}, archivePrefix={arXiv}, primaryClass={cs.SD}, url={https://arxiv.org/abs/2403.10380}, } """ _BIRDSET_DESCRIPTION = """\ Deep learning (DL) has greatly advanced audio classification, yet the field is limited by the scarcity of large-scale benchmark datasets that have propelled progress in other domains. While AudioSet is a pivotal step to bridge this gap as a universal-domain dataset, its restricted accessibility and limited range of evaluation use cases challenge its role as the sole resource. Therefore, we introduce BirdSet, a large-scale benchmark dataset for audio classification focusing on avian bioacoustics. BirdSet surpasses AudioSet with over 6,800 recording hours (+17%) from nearly 10,000 classes (x18) for training and more than 400 hours (x7) across eight strongly labeled evaluation datasets. It serves as a versatile resource for use cases such as multi-label classification, covariate shift or self-supervised learning. We benchmark six well-known DL models in multi-label classification across three distinct training scenarios and outline further evaluation use cases in audio classification. We host our dataset on Hugging Face for easy accessibility and offer an extensive codebase to reproduce our results. """ base_url = "https://huggingface.co/datasets/DBD-research-group/BirdSet/resolve/data" def _extract_all_to_same_folder(tar_path, output_dir): """custom extraction for tar.gz files, that extracts all files to output_dir without subfolders""" # check if data already exists if not os.path.isfile(output_dir) and os.path.isdir(output_dir) and os.listdir(output_dir): return output_dir os.makedirs(output_dir, exist_ok=True) with tarfile.open(tar_path, "r:gz") as tar: for member in tar.getmembers(): if member.isfile(): member.name = os.path.basename(member.name) tar.extract(member, path=output_dir) return output_dir def _extract_and_delete(dl_dir: dict, cache_dir: str = None) -> dict: """extracts downloaded files and deletes the archive file immediately, with progress bar. only the processed archive and its content are saved at the same time.""" audio_paths = {name: [] for name, data in dl_dir.items() if isinstance(data, list)} for name, data in dl_dir.items(): if not isinstance(data, list): continue # extract and immediately delete archives for path in tqdm(data, f"Extracting {name} split"): directory, filename = os.path.split(path) output_dir = os.path.join(cache_dir or directory, "extracted", filename.split(".")[0]) #audio_path = dl_manager.extract(path) # if all archive files are without subfolders this works just fine audio_path = _extract_all_to_same_folder(path, output_dir) os.remove(path) # datasets >3.0.0 hadels cach differently os.remove(f"{path}.lock") if os.path.exists(f"{path}.lock") else None os.remove(f"{path}.json") if os.path.exists(f"{path}.json") else None audio_paths[name].append(audio_path) return audio_paths class BirdSetConfig(datasets.BuilderConfig): def __init__( self, name, citation, class_list, genus_list, species_group_list, order_list, **kwargs): super().__init__(version=datasets.Version("0.0.4"), name=name, **kwargs) features = datasets.Features({ "audio": datasets.Audio(sampling_rate=32_000, mono=True, decode=False), "filepath": datasets.Value("string"), "start_time": datasets.Value("float64"), "end_time": datasets.Value("float64"), "low_freq": datasets.Value("int64"), "high_freq": datasets.Value("int64"), "ebird_code": datasets.ClassLabel(names=class_list), "ebird_code_multilabel": datasets.Sequence(datasets.ClassLabel(names=class_list)), "ebird_code_secondary": datasets.Sequence(datasets.Value("string")), "call_type": datasets.Value("string"), "sex": datasets.Value("string"), "lat": datasets.Value("float64"), "long": datasets.Value("float64"), "length": datasets.Value("int64"), "microphone": datasets.Value("string"), "license": datasets.Value("string"), "source": datasets.Value("string"), "local_time": datasets.Value("string"), "detected_events": datasets.Sequence(datasets.Sequence(datasets.Value("float64"))), "event_cluster": datasets.Sequence(datasets.Value("int64")), "peaks": datasets.Sequence(datasets.Value("float64")), "quality": datasets.Value("string"), "recordist": datasets.Value("string"), "genus": datasets.ClassLabel(names=genus_list), "species_group": datasets.ClassLabel(names=species_group_list), "order": datasets.ClassLabel(names=order_list), "genus_multilabel": datasets.Sequence(datasets.ClassLabel(names=genus_list)), "species_group_multilabel": datasets.Sequence(datasets.ClassLabel(names=species_group_list)), "order_multilabel": datasets.Sequence(datasets.ClassLabel(names=order_list)), }) self.features = features self.citation = citation class BirdSet(datasets.GeneratorBasedBuilder): """TODO: Short description of my dataset.""" # ram problems? DEFAULT_WRITER_BATCH_SIZE = 500 BUILDER_CONFIGS = [ BirdSetConfig( name="SSW", description=_SAPSUCKER_WOODS_DESCRIPTION, citation=_SAPSUCKER_WOODS_CITATION, data_dir=f"{base_url}/SSW", class_list=BIRD_NAMES_SAPSUCKER, genus_list=classes.GENUS_SSW, species_group_list=classes.SPECIES_GROUP_SSW, order_list=classes.ORDER_SSW, ), BirdSetConfig( name="SSW_xc", description=_SAPSUCKER_WOODS_DESCRIPTION, citation=_SAPSUCKER_WOODS_CITATION, data_dir=f"{base_url}/SSW", class_list=BIRD_NAMES_SAPSUCKER, genus_list=classes.GENUS_SSW, species_group_list=classes.SPECIES_GROUP_SSW, order_list=classes.ORDER_SSW, ), BirdSetConfig( name="SSW_scape", description=_SAPSUCKER_WOODS_DESCRIPTION, citation=_SAPSUCKER_WOODS_CITATION, data_dir=f"{base_url}/SSW", class_list=BIRD_NAMES_SAPSUCKER, genus_list=classes.GENUS_SSW, species_group_list=classes.SPECIES_GROUP_SSW, order_list=classes.ORDER_SSW, ), BirdSetConfig( name="PER", description=_AMAZON_BASIN_DESCRIPTION, citation=_AMAZON_BASIN_CITATION, data_dir=f"{base_url}/PER", class_list=BIRD_NAMES_AMAZON_BASIN, genus_list=classes.GENUS_PER, species_group_list=classes.SPECIES_GROUP_PER, order_list=classes.ORDER_PER, ), BirdSetConfig( name="PER_xc", description=_AMAZON_BASIN_DESCRIPTION, citation=_AMAZON_BASIN_CITATION, data_dir=f"{base_url}/PER", class_list=BIRD_NAMES_AMAZON_BASIN, genus_list=classes.GENUS_PER, species_group_list=classes.SPECIES_GROUP_PER, order_list=classes.ORDER_PER, ), BirdSetConfig( name="PER_scape", description=_AMAZON_BASIN_DESCRIPTION, citation=_AMAZON_BASIN_CITATION, data_dir=f"{base_url}/PER", class_list=BIRD_NAMES_AMAZON_BASIN, genus_list=classes.GENUS_PER, species_group_list=classes.SPECIES_GROUP_PER, order_list=classes.ORDER_PER, ), BirdSetConfig( name="UHH", description=_HAWAIIAN_ISLANDS_DESCRIPTION, citation=_HAWAIIAN_ISLANDS_CITATION, data_dir=f"{base_url}/UHH", class_list=BIRD_NAMES_HAWAII, genus_list=classes.GENUS_UHH, species_group_list=classes.SPECIES_GROUP_UHH, order_list=classes.ORDER_UHH, ), BirdSetConfig( name="UHH_xc", description=_HAWAIIAN_ISLANDS_DESCRIPTION, citation=_HAWAIIAN_ISLANDS_CITATION, data_dir=f"{base_url}/UHH", class_list=BIRD_NAMES_HAWAII, genus_list=classes.GENUS_UHH, species_group_list=classes.SPECIES_GROUP_UHH, order_list=classes.ORDER_UHH, ), BirdSetConfig( name="UHH_scape", description=_HAWAIIAN_ISLANDS_DESCRIPTION, citation=_HAWAIIAN_ISLANDS_CITATION, data_dir=f"{base_url}/UHH", class_list=BIRD_NAMES_HAWAII, genus_list=classes.GENUS_UHH, species_group_list=classes.SPECIES_GROUP_UHH, order_list=classes.ORDER_UHH, ), BirdSetConfig( name="SNE", description=_SIERRA_NEVADA_DESCRIPTION, citation=_SIERRA_NEVADA_CITATION, data_dir=f"{base_url}/SNE", class_list=BIRD_NAMES_SIERRA_NEVADA, genus_list=classes.GENUS_SNE, species_group_list=classes.SPECIES_GROUP_SNE, order_list=classes.ORDER_SNE, ), BirdSetConfig( name="SNE_xc", description=_SIERRA_NEVADA_DESCRIPTION, citation=_SIERRA_NEVADA_CITATION, data_dir=f"{base_url}/SNE", class_list=BIRD_NAMES_SIERRA_NEVADA, genus_list=classes.GENUS_SNE, species_group_list=classes.SPECIES_GROUP_SNE, order_list=classes.ORDER_SNE, ), BirdSetConfig( name="SNE_scape", description=_SIERRA_NEVADA_DESCRIPTION, citation=_SIERRA_NEVADA_CITATION, data_dir=f"{base_url}/SNE", class_list=BIRD_NAMES_SIERRA_NEVADA, genus_list=classes.GENUS_SNE, species_group_list=classes.SPECIES_GROUP_SNE, order_list=classes.ORDER_SNE, ), BirdSetConfig( name="POW", description=_POWDERMILL_NATURE_DESCRIPTION, citation=_POWDERMILL_NATURE_CITATION, data_dir=f"{base_url}/POW", class_list=BIRD_NAMES_POWDERMILL_NATURE, genus_list=classes.GENUS_POW, species_group_list=classes.SPECIES_GROUP_POW, order_list=classes.ORDER_POW, ), BirdSetConfig( name="POW_xc", description=_POWDERMILL_NATURE_DESCRIPTION, citation=_POWDERMILL_NATURE_CITATION, data_dir=f"{base_url}/POW", class_list=BIRD_NAMES_POWDERMILL_NATURE, genus_list=classes.GENUS_POW, species_group_list=classes.SPECIES_GROUP_POW, order_list=classes.ORDER_POW, ), BirdSetConfig( name="POW_scape", description=_POWDERMILL_NATURE_DESCRIPTION, citation=_POWDERMILL_NATURE_CITATION, data_dir=f"{base_url}/POW", class_list=BIRD_NAMES_POWDERMILL_NATURE, genus_list=classes.GENUS_POW, species_group_list=classes.SPECIES_GROUP_POW, order_list=classes.ORDER_POW, ), BirdSetConfig( name="HSN", description=_HIGH_SIERRAS_DESCRIPTION, citation=_HIGH_SIERRAS_CITATION, data_dir=f"{base_url}/HSN", class_list=BIRD_NAMES_HIGH_SIERRAS, genus_list=classes.GENUS_HSN, species_group_list=classes.SPECIES_GROUP_HSN, order_list=classes.ORDER_HSN, ), BirdSetConfig( name="HSN_xc", description=_HIGH_SIERRAS_DESCRIPTION, citation=_HIGH_SIERRAS_CITATION, data_dir=f"{base_url}/HSN", class_list=BIRD_NAMES_HIGH_SIERRAS, genus_list=classes.GENUS_HSN, species_group_list=classes.SPECIES_GROUP_HSN, order_list=classes.ORDER_HSN, ), BirdSetConfig( name="HSN_scape", description=_HIGH_SIERRAS_DESCRIPTION, citation=_HIGH_SIERRAS_CITATION, data_dir=f"{base_url}/HSN", class_list=BIRD_NAMES_HIGH_SIERRAS, genus_list=classes.GENUS_HSN, species_group_list=classes.SPECIES_GROUP_HSN, order_list=classes.ORDER_HSN, ), BirdSetConfig( name="NES", description=_COLUMBIA_COSTA_RICA_DESCRIPTION, citation=_COLUMBIA_COSTA_RICA_CITATION, data_dir=f"{base_url}/NES", class_list=BIRD_NAMES_COLUMBIA_COSTA_RICA, genus_list=classes.GENUS_NES, species_group_list=classes.SPECIES_GROUP_NES, order_list=classes.ORDER_NES, ), BirdSetConfig( name="NES_xc", description=_COLUMBIA_COSTA_RICA_DESCRIPTION, citation=_COLUMBIA_COSTA_RICA_CITATION, data_dir=f"{base_url}/NES", class_list=BIRD_NAMES_COLUMBIA_COSTA_RICA, genus_list=classes.GENUS_NES, species_group_list=classes.SPECIES_GROUP_NES, order_list=classes.ORDER_NES, ), BirdSetConfig( name="NES_scape", description=_COLUMBIA_COSTA_RICA_DESCRIPTION, citation=_COLUMBIA_COSTA_RICA_CITATION, data_dir=f"{base_url}/NES", class_list=BIRD_NAMES_COLUMBIA_COSTA_RICA, genus_list=classes.GENUS_NES, species_group_list=classes.SPECIES_GROUP_NES, order_list=classes.ORDER_NES, ), BirdSetConfig( name="NBP", description=_NIPS4BPLUS_DESCRIPTION, citation=_NIPS4BPLUS_CITATION, data_dir=f"{base_url}/NBP", class_list=BIRD_NAMES_NIPS4BPLUS, genus_list=classes.GENUS_NBP, species_group_list=classes.SPECIES_GROUP_NBP, order_list=classes.ORDER_NBP, ), BirdSetConfig( name="NBP_xc", description=_NIPS4BPLUS_DESCRIPTION, citation=_NIPS4BPLUS_CITATION, data_dir=f"{base_url}/NBP", class_list=BIRD_NAMES_NIPS4BPLUS, genus_list=classes.GENUS_NBP, species_group_list=classes.SPECIES_GROUP_NBP, order_list=classes.ORDER_NBP, ), BirdSetConfig( name="NBP_scape", description=_NIPS4BPLUS_DESCRIPTION, citation=_NIPS4BPLUS_CITATION, data_dir=f"{base_url}/NBP", class_list=BIRD_NAMES_NIPS4BPLUS, genus_list=classes.GENUS_NBP, species_group_list=classes.SPECIES_GROUP_NBP, order_list=classes.ORDER_NBP, ), BirdSetConfig( name="XCM", description="TODO", citation="TODO", data_dir=f"{base_url}/XCM", class_list=BIRD_NAMES_XENOCANTO_M, genus_list=classes.GENUS_XCM, species_group_list=classes.SPECIES_GROUP_XCM, order_list=classes.ORDER_XCM, ), BirdSetConfig( name="XCL", description="TODO", citation="TODO", data_dir=f"{base_url}/XCL", class_list=BIRD_NAMES_XENOCANTO, genus_list=classes.GENUS_XCL, species_group_list=classes.SPECIES_GROUP_XCL, order_list=classes.ORDER_XCL, ), ] def _info(self): return datasets.DatasetInfo( description=_BIRDSET_DESCRIPTION + self.config.description, features=self.config.features, citation=self.config.citation + "\n" + _BIRDSET_CITATION, ) def _split_generators(self, dl_manager): ds_name = self.config.name # settings for how much archives (tar.gz) files are uploaded for a specific dataset train_files = {"PER": 11, "NES": 13, "UHH": 5, "HSN": 7, "NBP": 32, "POW": 9, "SSW": 29, "SNE": 21, "XCM": 182, "XCL": 98} test_files = {"PER": 3, "NES": 8, "UHH": 7, "HSN": 3, "NBP": 1, "POW": 3, "SSW": 36, "SNE": 5} test_5s_files = {"PER": 1, "NES": 1, "UHH": 1, "HSN": 1, "NBP": 1, "POW": 1, "SSW": 4, "SNE": 1} # different configs, determine what needs to be downloaded if self.config.name.endswith("_xc"): ds_name = ds_name[:-3] dl_dir = dl_manager.download({ "train": [os.path.join(self.config.data_dir, f"{ds_name}_train_shard_{n:04d}.tar.gz") for n in range(1, train_files[ds_name] + 1)], "meta_train": os.path.join(self.config.data_dir, f"{ds_name}_metadata_train.parquet"), }) elif self.config.name.endswith("_scape"): ds_name = ds_name[:-6] dl_dir = dl_manager.download({ "test": [os.path.join(self.config.data_dir, f"{ds_name}_test_shard_{n:04d}.tar.gz") for n in range(1, test_files[ds_name] + 1)], "test_5s": [os.path.join(self.config.data_dir, f"{ds_name}_test5s_shard_{n:04d}.tar.gz") for n in range(1, test_5s_files[ds_name] + 1)], "meta_test": os.path.join(self.config.data_dir, f"{ds_name}_metadata_test.parquet"), "meta_test_5s": os.path.join(self.config.data_dir, f"{ds_name}_metadata_test_5s.parquet"), }) # use POW for XCM/XCL validation elif self.config.name.startswith("XC"): dl_dir = dl_manager.download({ "train": [os.path.join(self.config.data_dir, f"{ds_name}_shard_{n:04d}.tar.gz") for n in range(1, train_files[ds_name] + 1)], "valid": [os.path.join(self.config.data_dir[:-3] + "POW", f"POW_test5s_shard_{n:04d}.tar.gz") for n in range(1, test_5s_files["POW"] + 1)], "meta_train": os.path.join(self.config.data_dir, f"{ds_name}_metadata.parquet"), "meta_valid": os.path.join(self.config.data_dir[:-3] + "POW", f"POW_metadata_test_5s.parquet"), }) else: dl_dir = dl_manager.download({ "train": [os.path.join(self.config.data_dir, f"{ds_name}_train_shard_{n:04d}.tar.gz") for n in range(1, train_files[ds_name] + 1)], "test": [os.path.join(self.config.data_dir, f"{ds_name}_test_shard_{n:04d}.tar.gz") for n in range(1, test_files[ds_name] + 1)], "test_5s": [os.path.join(self.config.data_dir, f"{ds_name}_test5s_shard_{n:04d}.tar.gz") for n in range(1, test_5s_files[ds_name] + 1)], "meta_train": os.path.join(self.config.data_dir, f"{ds_name}_metadata_train.parquet"), "meta_test": os.path.join(self.config.data_dir, f"{ds_name}_metadata_test.parquet"), "meta_test_5s": os.path.join(self.config.data_dir, f"{ds_name}_metadata_test_5s.parquet"), }) # custom extraction that deletes archives right after extraction audio_paths = _extract_and_delete(dl_dir, dl_manager.download_config.cache_dir) if not dl_manager.is_streaming else None # construct split generators # assumes every key in dl_dir of NAME also has meta_NAME names = [name for name in dl_dir.keys() if not name.startswith("meta_")] is_streaming = dl_manager.is_streaming return [datasets.SplitGenerator( name=name, gen_kwargs={ "audio_archive_iterators": (dl_manager.iter_archive(archive_path) for archive_path in dl_dir[name]) if is_streaming else () , "audio_extracted_paths": audio_paths[name] if not is_streaming else (), "meta_path": dl_dir[f"meta_{name}"], "split": name } ) for name in names] def _generate_examples(self, audio_archive_iterators, audio_extracted_paths, meta_path, split): metadata = pd.read_parquet(meta_path) if metadata.index.name != "filepath": metadata.index = metadata["filepath"].str.split("/").apply(lambda x: x[-1]) idx = 0 # in case of streaming for audio_archive_iterator in audio_archive_iterators: for audio_path_in_archive, audio_file in audio_archive_iterator: file_name = os.path.split(audio_path_in_archive)[-1] rows = metadata.loc[[file_name]] audio = audio_file.read() for _, row in rows.iterrows(): yield idx, self._metadata_from_row(row, split, audio_path=file_name, audio=audio) idx += 1 # in case of not streaming for audio_extracted_path in audio_extracted_paths: audio_files = os.listdir(audio_extracted_path) current_metadata = metadata.loc[audio_files] for audio_file, row in current_metadata.iterrows(): audio_path = os.path.join(audio_extracted_path, audio_file) yield idx, self._metadata_from_row(row, split, audio_path=audio_path) idx += 1 @staticmethod def _metadata_from_row(row, split: str, audio_path=None, audio=None) -> dict: return {"audio": audio_path if not audio else {"path": None, "bytes": audio}, "filepath": audio_path, "start_time": row["start_time"], "end_time": row["end_time"], "low_freq": row["low_freq"], "high_freq": row["high_freq"], "ebird_code": row["ebird_code"] if split != "test_5s" else None, "ebird_code_multilabel": row.get("ebird_code_multilabel", None), "ebird_code_secondary": row.get("ebird_code_secondary", None), "call_type": row["call_type"], "sex": row["sex"], "lat": row["lat"], "long": row["long"], "length": row.get("length", None), "microphone": row["microphone"], "license": row.get("license", None), "source": row["source"], "local_time": row["local_time"], "detected_events": row.get("detected_events", None), "event_cluster": row.get("event_cluster", None), "peaks": row.get("peaks", None), "quality": row.get("quality", None), "recordist": row.get("recordist", None), "genus": row.get("genus", None) if split != "test_5s" else None, "species_group": row.get("species_group", None) if split != "test_5s" else None, "order": row.get("order", None) if split != "test_5s" else None, "genus_multilabel": row.get("genus_multilabel", [row.get("genus")]), "species_group_multilabel": row.get("species_group_multilabel", [row.get("species_group")]), "order_multilabel": row.get("order_multilabel", [row.get("order")]), }